注意力机制和特征融合的自动抠图算法  

Automatic Image Matting with Attention Mechanism and Feature Fusion

作　　者：王欣 王琦琦 杨国威 郭肖勇 Wang Xin;Wang Qiqi;Yang Guowei;Guo Xiaoyong(College of Electronic Information and Automation,Tianjin University of Science&Technology,Tianjin 300222)

机构地区：[1]天津科技大学电子信息与自动化学院,天津300222

出　　处：《计算机辅助设计与图形学学报》2020年第9期1473-1483,共11页Journal of Computer-Aided Design & Computer Graphics

摘　　要：针对目前人工抠图工作量大,而自动抠图无法区分多个实例的问题,提出了一种注意力机制和特征融合的自动抠图算法.该算法由预分割模块和Alpha抠图模块2部分组成,分别采用了不同的网络结构.其中预分割模块是使用迁移学习方法对Mask Scoring R-CNN进行微调实现了对多实例自然图像的实例分割,从而得到输入图像前景个体的二值化分割图.而Alpha抠图模块在此基础上首先对二值化分割图预处理生成三分图,然后将三分图与原输入图像一起输入Alpha抠图模块网络.通过为Alpha抠图模块设计不同的解码策略和注意力机制,实现了对图像细节信息的精确恢复.在后续对自制车辆数据集的无人工交互前景车辆Alpha估计对比实验中,相比现有DIM算法,该算法的SAD降低19.2%,MSE降低26.3%,达到了更高的抠图精度.In response to the current problem of the excessive workload of manual matting and the inability of automatic matting to distinguish between multiple instances,an automatic matting algorithm with attention mechanism and feature fusion is proposed.The algorithm consists of two parts,the pre-segmentation module and the Alpha matting module,which adopt different structures respectively.The presegmentation module uses transfer learning method to fine-tune a mask scoring R-CNN to implement instance segmentation of multi-instance natural images and obtain a binary segmentation of the foreground individuals.Based on this,the Alpha module first pre-processes the binary segmentation map into a generate trimap,which is then fed into the Alpha matting module network along with the original input image.By designing different decoding strategies and attention mechanisms for the Alpha matting module,the accurate recovery of input image details is achieved.In a follow-up comparison experiment of foreground vehicle Alpha estimation,which uses a homemade vehicle data set but without human interaction,this algorithm achieves a higher matting accuracy with 19.2%lower SAD and 26.3%lower MSE than the existing DIM algorithm.

关 键 词：实例分割 注意力机制 特征融合 Alpha抠图算法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]